The public conversation about AI splits into two tired genres. One is the breathless product launch: this changes everything, the old world is over, here is an embeddable widget. The other is the alarm bell: it will take your job, your children’s future, your last shred of meaning. Both are exhausting. Both are mostly noise.
Between those poles, something more useful has been happening — quietly, without a manifesto. People are learning to work with these systems. Not to be replaced by them, not to worship them. To deploy them as tools that accelerate thinking without outsourcing it. That distinction matters more than any benchmark number.
The best metaphor I have found is a colleague. An odd one.
Imagine someone who has read every book, every paper, every README. Their recall is near-perfect. They never get bored, never need sleep, and respond to every question — no matter how trivial — with a full paragraph. But they have never inhabited a body, never faced a consequence, never had to sit with a decision that couldn’t be rolled back. They are brilliant and entirely ungrounded.
This is the shape of the thing. The skill in working with it is not prompt engineering. It is knowing when to listen and when to overrule.
I see two failure modes. The first is deference: treating the AI as an oracle, accepting its confident prose as fact, letting it set the agenda. This is how you end up with hallucinated packages, plausible-sounding nonsense, and a slow erosion of your own judgement. The second is dismissal: refusing to engage because it made a mistake once, or because the hype is annoying. This is how you leave a genuine lever on the table.
The sweet spot is narrower than either.
Some things these systems are genuinely good at, right now:
Summarizing and restructuring. Give them a mess — meeting notes, a long thread, a spec document that three people edited in conflicting directions — and they will hand back something coherent. Not perfect, but coherent enough that you can start editing instead of starting from scratch.
First drafts. Not final drafts. The difference between writing something from a blank page and revising something bad-but-present is enormous. AI collapses that gap.
Exploring the shape of a problem. “Here is what I’m trying to do. What am I missing?” It will surface angles you hadn’t considered — not because it is wise, but because it has seen ten thousand similar problems and remembers the common failure points.
Translating between domains. Turning a SQL query into an explanation your PM can read. Converting a legal clause into plain English. Generating the boilerplate so you can focus on the part that actually requires thought.
Rubber-ducking with recall. Talking through a decision with someone who can pull up relevant research, counterarguments, and edge cases on demand — without needing to be brought up to speed.
Now the things they are still bad at:
Knowing when they are wrong. An AI will defend a false statement with the same syntactic confidence it brings to a true one. It does not have uncertainty; it has probability distributions dressed up as prose. You must verify.
Understanding consequence. It can describe the legal implications of a contract clause, but it does not feel the weight of a bad deal. That weight is information, and it is missing.
Taste. It can mimic style. It cannot originate one that comes from having lived through a particular set of experiences. The difference between a good sentence and a great one is often incompressible — it’s why a particular word lands, why a structure feels right. That judgment is yours.
Long-range coherence. The longer the output, the more it drifts. A paragraph is tight; a chapter is mush. This improves with each model generation but the fundamental problem — no persistent internal state that maps to intention — means you still need a human holding the arc.
So how do you actually use one?
Rule one: you are the bottleneck, not the AI. Ask it to do the part you are slow at. If you write fast but research slowly, have it research. If you can design a schema in your head but hate typing out migrations, have it generate the boilerplate. The goal is not to offload thinking — it’s to offload the parts of the work that sit between thoughts.
Rule two: always make it produce something you can check. A summary can be checked against the original. Generated code can be run. A translation can be read by a native speaker. If the output is unfalsifiable — “what is the meaning of life” — you are not working, you are entertaining yourself.
Rule three: iterate. The first answer is the shallowest. Push back: “that’s too vague,” “what’s the counterargument,” “give me three alternatives and rank them.” The system will comply, and the second or third pass is almost always sharper than the first.
Rule four: keep the hard problems for yourself. If a decision requires judgment, values, or accepting an irreversible trade-off, the AI should inform it, not make it. The moment you delegate those, you have stopped being the operator and started being the passenger.
None of this is permanent. The models will get better, the failure modes will shift, and some of what I just wrote will be wrong within a year. But the core posture — engaged but not deferential, curious but not credulous — ages well regardless of the technology underneath it.
The people who thrive in the next decade will not be the ones who learned the most prompt tricks. They will be the ones who learned to sit next to a machine that is faster than them at most things, and still know when to say no, that’s not quite right — let me think about this one myself.